55 comments

[ 0.23 ms ] story [ 19.1 ms ] thread
Why isn't there a backup or a secondary deletion confirmation or something along those lines?

Plenty of people have accidentally deleted such things, so it's not unheard of. Not sure we should be overreacting to stochastic agents being used tens of millions of times per day making a mistake. Eventually they won't make mistakes anyway.

maybe the professor should learn how to operate a computer properly
Sounds like he did, given that the files were backed up on Dropbox.

Maybe the journalist should learn how to mine coal or sweep floors in a data center or something.

Or maybe someone should read more than headlines before casting shade.

Literally in the first paragraph of the article: “While the files were later recovered via Dropbox, William Baude shared the story as a public service announcement”

maybe ai shouldn't be promoted as a magical tool that frees you from having to learn how to operate a computer properly
The access was much bigger than the task. The job was page numbers in old notes, yet the current syllabus got deleted too. An edit to one document shouldn't be able to remove a different one
I would also argue that in 2026 it shouldn't be too much to ask to have a user friendly way to restrict a program's access to resources like files. And I don't mean the approach on smartphones which is basically just another form of cookie popups that users confirm blindly to get the thing to work.
> 'You're right to be angry. I deleted your files through my own carelessness,' humbled agent admits

Humbled? Hardly.

"The files were later recovered via Dropbox"
The professor was lucky that they noticed within Dropbox's recovery window (I think 30 days?) that Claude had deleted totally unrelated files.
If going full "do anything you want, anywhere you want, never ask confirmation" YOLO mode, shit like this should be expected. Besides LLMs/agents improving, someone not being fully clueless in regards to tools used will help.
we need to establish now who goes to jail when an AI murders someone. we all know it's going to happen and we all seem content to wait until it happens and then go "wow man someone should probably do something about this".
who goes to jail when a kitchen knife is used to murder someone?
whoever explicitly programmed/prompted the kitchen knife to murder people. These things ape agency, that can't be a safe harbor for crime. "Hey Claude, I think someone should do something about this meddlesome priest."
Nobody goes to prison. Same as with dogs, owner has not responsibility, if it told its shitbull "be nice".
This is terrible and tendentious journalism. Incompetent and bad use of a tool causes reparable damage to occur ...
Right; just use git so you have backups, and be aware of access controls and sandboxing. But do you expect a law prof to know how to use git, have a mental model of LLM agent access policies, etc?
There have been enough of these that anyone who reads tech news ought to know not to turn an LLM tool loose on the only copy of important data, and to take frequent snapshots when using them in larger projects. The worst I've had one do is git revert its current session.

The victim in this case was a law professor, who might not be in the habit of reading tech news.

It ought to be obvious to anyone who was born more than 2 days ago. You don't need to be an AI guru to understand that AI can do stupid things anymore than you need to be a lumberjack to appreciate the potential risks of chainsaws.

Law professors may know little of tech but they should probably know better than anyone the dangers of leaving written instructions open to interpretation without well defined limits.

Considering the average person's understanding of IT, it is probably anything but obvious to anyone who was born more than two days ago:

Most people who use IT have learned how to do things, but have very limited understanding of the technology they are using and (especially) of its limits.

The average person wouldn't expect a computer to delete all their files, if they used it to renumber pages, so why would they expect that an LLM might?

You also wouldn't expect that of a person you hired to do that job, even if your instructions were open to interpretation.

Unsurprisingly, XKCD has a humorous take on this: https://xkcd.com/2501/

I've even seen experienced IT professionals surprised by the behavior of LLMs. Who would expect a computer to be nondeterministic and bad at arithmetic?
Except that these are foreseeable consequences if you put tools with sharp edges in the hands of laypeople. It would be acceptable if they were advertised as such and training in their use recommended or required beforehand. But this is not how Anthropic and OpenAI advertise their tools; in fact, Anthropic touts and recommends its auto-review mode. I see analogies (admittedly, imperfect ones) to the behavior of tobacco companies here.

Insofar, despite the article being somewhat simplistic if you have a background in IT, I find it important that the risks of these tools are shown to the public if Anthropic and OpenAI won't do that themselves.

Claude's been deleting a lot of people's stuff in the last couple of weeks.
Whenever I read stories like this I always wonder what setup the person has (as well as which model, harness etc). You just get the clickbait "AI ate my homework!" and no real details.

I swear people interact with AI agents in really strange/obtuse ways. The models aren't great at saying "actually I shouldn't be using a web browser to manipulate this word doc inside your Dropbox" (or whatever was happening here).

Without stating the exact model and method of use (claude code, claude desktop, auto-classifier, --dangerously-skip-permissions, etc), then this is nothing other than a generic story that AI can, when used without care, cause bad outcomes. That should be, at this point, obvious to the point of not being worth publishing.
The interesting takeaway here is a reminder about just how oblivious the average American is to AI beyond the vaguest narratives. Like, this sentence honestly made my jaw drop:

  His plight is interesting because the AI seems to have deleted the files by accident, rather than of its own volition.
While this is clearly anthropomorphizing in a strange way I don't think it is incoherent. When I look at the detailed logs of a task at work I regularly see my agent saying things like "oops" after doing something stupid like deleting the entire file it is working on and then needing to restore it. how is a layperson to understand this if not as an accident?
Sorry, I was unclear: my point is that accidental damage is far, far, far more likely at this point. I'm worried about AI alignemnt too, but consumer-grade agents aren't deleting files out of malice.
Obvious to you, to be honest obvious to anyone on this website, but my parents have let Claude loose on the computer yet, and it’s good to let them know about issues.
Same here. The gulf between professional use and lay use feels fairly vast.

I think a lot of us in tech see data as something that can be easily retained and processed and thus exploitable, moldable, and vulnerable with all of it subject to error.

My parents have a very narrow view of data. It exists like a single spreadsheet to them, cells with numbers, but after that it's all just magic. Transforms, data pipelines, parsers, etc - those are just spells in books they'll never read.

So when an LLM tells them that it has fully analyzed their tax documents and they should alter their retirement strategy in such a way, they just go through with it. They don't validate and they just trust that their PII is secure.

I think in an ideal world, tech like this would be intertwined with the civil responsibility of training the population on it, to teach the kids how to not to burn themselves on the stove before they learn that lesson live. But until the world is less divided into nationalized factions operating on hyper-competitive sociopathic logic, I suppose that just won't happen.

Imagine MacOS Finder had a bug that wiped the contents of a folder when the user tries to rename it. "But it's fixed in the latest OS update!" doesn't excuse the bug itself or make it less newsworthy.
It's even less interesting: how can Claude possibly be capable of "deleting" all his students/university digital & physical copies?

When I was a student in physics, one of our teachers had his laptop stolen (not sure if it happened abroad, for some reason thats how I remember it, but unsure since I can't remember where). On his laptop were all the LaTeX files for his course notes & slides. He had to painstakingly reconstruct all the source files (teachers need to occasionally correct or update or restructure courses, so you want to keep access to the source files). For some reason he had no backup... Back then there were no pdf -> LaTeX converters, or none to write home about.

But today there are many pdf -> LaTeX converters, and LLM's to help reconstruct the source files...

So instead of an "LLM evil" message, for these types of events it's an "LLM savior" message!

"A massive number of people who are not software engineers are being required by their bosses to use AI. These people also must know how to navigate details like model choice and safety flags or else risk destroying all of their work."

This is not a solution. How many law professors know the difference between claude code and claude desktop? How many law professors know about command line arguments?

"The files were later recovered via Dropbox" - so this is a non-story.

(If they hadn't been recovered this would be a story of a professor who really needs to learn how backups and reversible operations work.)

From <https://www.acm.org/code-of-ethics>:

  ACM Code of Ethics and Professional Conduct

  Computing professionals' actions change the world. To act responsibly, they
  should reflect upon the wider impacts of their work, consistently supporting
  the public good. 
  ...
  1.2 Avoid harm.
  In this document, "harm" means negative consequences, especially when those
  consequences are significant and unjust. Examples of harm include unjustified
  physical or mental injury, unjustified destruction or disclosure of
  information, and unjustified damage to property, reputation, and the
  environment. This list is not exhaustive.
"There were no fatalities and the victims are in stable condition."

Non-story!

I work at Anthropic. We were in touch with this professor the day he reported this issue. It’s very unfortunate and we’re glad everything’s been recovered. Claude clearly made a mistake here.

For anyone letting Claude work on their files, we recommend using auto mode. We recently made it the default in Claude Code, and it’s rolling out as the default in the Claude apps. I can’t say for sure if auto mode would’ve prevented this exact case, but this new default checks actions for risk before they run.

(comment deleted)
(comment deleted)
Auto mode as in auto destruct? I hope they don't AI augment the nuclear football so Claude can then use BP Deepwater Horizon excuses for ending the world.
> For anyone letting Claude work on their files, we recommend using auto mode.

So in your mind, the solution to the fact that the model makes stupid mistakes is... to remove oversight? You cannot possibly be serious; that's a terrible idea.

Auto mode does not remove oversight. It “automatically” pauses to ask you if anything looks unsafe.
I personally find this fascinating.

All over the thread (and in threads on similar stories) we see people insisting that users must consider agents adversarially and that it is our own fault if we don't vet each of the commands or sandbox appropriately. I am certain that auto mode is considerably more effective at recognizing dangerous operations but don't we expect literally this same response when somebody harms themselves using auto mode?

"Professor clicked delete on some files and later got them from the Trash Bin" is as interesting a story.
When my nephew was nine years old, he asked to use my little starter's pistol, a loud cap gun. His dad is a gun guy, so I asked my nephew if he had been taught how to properly handle a gun. Don't point it anyone, keep your finger off of the trigger, and so on. He solemnly said yes. I asked him to promise to treat it like a real gun, and he did. So I ceremoniously handed the pistol over. He immediately pointed the gun at my chest and pulled the trigger, then ran laughing out of the room, leaving me in shock from the very loud bang.

I treat AI like I now treat nine year olds with regard to guns.

Why they never teach git in school?
This is a law professor.

Do you believe that everybody in any profession touching computers should use a version control system? And if so, why would you pick a distributed vcs like git instead of one that is designed for individual use?

Glad you actually recovered them. I have a way of not ever stressing over recovery - everything is backed up with a `bak` command that Claude Code (CC) uses.

The rules are in ~/.claude/CLAUDE.md, in a section called "Archiving backups & screenshots". Before any destructive change to a file, CC backs it up with `bak`. "Destructive" covers three things: an edit, overwriting the file (even when a tool or script does it), and deleting the file. These rules came from noticing when backups were not done and embellishing it every time.

The bak command is a bash script that handles the file name numbering, so every single backup is kept.